For much of the artificial-intelligence boom, the conversation has centered on capital and compute. How many GPUs can be acquired? How quickly can another data center be financed and built? How much are the hyperscalers spending? How many gigawatts of new campuses have been announced?
Increasingly, another question sits underneath all of them: Where does the electricity come from?
The answer is beginning to determine which AI ambitions become physical infrastructure and which remain plans on paper.
The International Energy Agency estimated that U.S. data centers consumed roughly 180 terawatt-hours of electricity in 2024. By 2030, it expects that figure to increase by approximately another 240 TWh—an increase of about 130 percent. The IEA now expects data centers to account for roughly half of the growth in U.S. electricity demand through the end of the decade. IEA
Those numbers change the nature of the AI infrastructure problem.
Capital Is Mobile. Electricity Is Physical.
Money can cross a continent electronically. Capital can be redirected from one project to another in minutes. GPUs can be purchased internationally and transported to wherever someone is prepared to install them.
A megawatt is different.
Electricity has to be generated somewhere. Generation equipment has to exist. Fuel has to reach it. Transmission infrastructure has to carry the electricity. Substations and transformers have to accommodate the load. A data center has to secure an interconnection, and the local electrical system has to be capable of supporting that enormous new concentration of demand.
That physicality introduces something the software economy has often been able to avoid: hard geographic constraints.
The scale is already substantial. The IEA notes that a conventional data center might require 10–25 MW, while an AI-focused hyperscale facility can require 100 MW or more. Facilities under construction are reaching into the gigawatts, and proposed campuses can be larger still. IEA
At those scales, a data center is no longer simply another commercial building requesting electrical service. It begins to resemble heavy industry.
The AI Economy Is Meeting the Energy Economy
This is where some of the enormous numbers surrounding the AI boom need to be interpreted carefully.
The United States currently has a huge pipeline of announced data-center capacity. One August 2026 project database counted 318.6 GW of planned or under-construction capacity, including 90.4 GW already under construction. Those figures should not be read as 318.6 GW of imminent operating load—the underlying projects are at different stages and their reported power figures are not perfectly standardized—but the scale of the pipeline illustrates the magnitude of the electricity being sought. US Data Center Map
The grid cannot respond to that kind of demand with software-industry speed.
Google has publicly identified transmission as one of the largest obstacles to connecting new U.S. data centers, with waits in some places reaching as long as twelve years. Large gas turbines have also developed extraordinary lead times; industry reporting in 2026 put the average around five years, with some stretching to seven. Reuters
PJM offers an even more immediate illustration. Its July 2026 capacity auction procured 138,318 MW of generation and demand-response resources, yet still came in 6,831 MW below its reliability requirement. The auction price hit the federally approved cap. PJM serves a region containing some of the country's most important data-center markets. PJM Inside Lines
We therefore have a peculiar economic situation emerging. There can be sufficient investor appetite for another AI campus, sufficient demand for its computational output, sufficient land on which to construct it, and perhaps even access to the chips required to fill it—while the project still waits on electrons.
Capital can be committed years before power can be delivered.
The Watt Becomes an Economic Input to Intelligence
This is the part that deserves more attention.
A watt is obviously not intelligence. Electricity by itself performs no inference, trains no model and produces no useful cognition. Yet every computational operation ultimately requires energy.
That creates a physical chain:
Energy → Electrical Capacity → Compute → Machine Intelligence → Economic Output
The greater the economic value we derive from machine intelligence, the more strategically important the upstream inputs into that chain become.
This is why I have argued throughout Power → Compute → Intelligence that energy policy can no longer be treated as something separate from technology policy. A country can possess extraordinary software talent, abundant investment capital and world-leading semiconductor technology while still constraining the productive deployment of those assets if it cannot expand electrical capacity rapidly enough.
The constraint simply moves upstream.
For decades, much of the digital economy trained us to think that computing resources were effectively detached from geography. Cloud computing reinforced that abstraction. A developer could request another server instance without knowing—or particularly caring—which power plant ultimately supplied the electricity behind it.
Artificial intelligence is bringing the physical layer back into view.
At sufficiently large scale, compute is industrial infrastructure.
And industrial infrastructure consumes power.
The Constraint Is Already Changing Behavior
Markets respond to constraints by trying to route around them.
That is exactly what is beginning to happen.
Data-center developers and technology companies are increasingly exploring behind-the-meter generation, colocated generation, dedicated natural-gas facilities, nuclear power agreements, advanced geothermal, energy storage and other arrangements that reduce dependence on conventional grid interconnection schedules. The IEA reports that natural gas currently provides more than 40 percent of U.S. data-center electricity and projects it to supply more than 130 TWh of additional annual generation for the sector through 2030. Nuclear is also receiving renewed attention as operators seek large quantities of firm power. IEA
This is an important signal.
Companies whose core businesses are software, cloud services and artificial intelligence are being pulled progressively deeper into the economics of electricity generation.
That should tell us something about where the bottleneck has moved.
The most valuable AI infrastructure may increasingly be located where several things converge: land, fiber, capital, favorable regulation, cooling capability—and above all, large quantities of reliable power that can actually be delivered on schedule.
The economic-development implications are considerable. States and regions accustomed to competing for technology investment through tax incentives, available real estate and workforce programs may find that dispatchable megawatts and interconnection timelines become just as important. An industrial site with power available now may be more strategically valuable than a seemingly superior site where sufficient electricity will not arrive for another six years.
Watts Are the Ceiling—But They Also Raise the Next Question
There is an important limit to this argument. Efficiency continues improving. The IEA reported in 2026 that electricity consumption per AI task has been falling rapidly. Better accelerators, model architectures, cooling systems and software optimization can produce more computation from a given quantity of electricity. IEA
That is enormously important.
Yet efficiency does not eliminate the power constraint when total demand for computation grows faster than efficiency reduces energy per operation. The IEA still expects data-center electricity consumption to double globally by 2030, with AI-focused facilities growing considerably faster. IEA
The relevant economic measure therefore becomes something closer to productive intelligence per watt.
How much useful computation can an economy extract from the electrical capacity available to it?
That question also leads directly into the next stage of this series.
If electrical capacity becomes scarce, expensive or geographically constrained, then it is no longer sufficient to ask how much compute we should build. We also have to ask where computation should happen.
Frontier-model training may justify enormous concentrations of power and accelerators in hyperscale facilities. Other forms of inference may belong much closer to the physical processes they serve: on a farm managing irrigation and cold storage, inside a factory interpreting machine telemetry, at a mine coordinating equipment, within an electrical system balancing loads, or at a logistics facility processing cameras and sensors locally.
The watt may be the constraint.
Where we turn those watts into intelligence becomes the next question.
Analysis 10: Intelligence Has a Geography.